Prototype-Guided Diffusion for Digital Pathology: Achieving Foundation Model Performance with Minimal Clinical Data

TL;DR

The paper tackles the data demands of foundation models in digital pathology by introducing prototype-guided diffusion to generate high-fidelity synthetic histopathology data. It trains a latent diffusion model with classifier guidance using unsupervised histological prototypes and evaluates the resulting SSL features via ABMIL on downstream cancer subtyping and survival tasks. The approach achieves competitive performance with orders of magnitude less real data, and a hybrid synthetic+real dataset sets new benchmarks across several tasks, including NSCLC survival and prostate cancer BCR. Notably, the authors report a digital pathology foundation model trained on $1.7\times 10^6$ synthetic images achieving performance comparable to models trained on real datasets roughly $60\times$ larger, underscoring the potential to reduce reliance on extensive clinical data while accelerating model development.

Abstract

Foundation models in digital pathology use massive datasets to learn useful compact feature representations of complex histology images. However, there is limited transparency into what drives the correlation between dataset size and performance, raising the question of whether simply adding more data to increase performance is always necessary. In this study, we propose a prototype-guided diffusion model to generate high-fidelity synthetic pathology data at scale, enabling large-scale self-supervised learning and reducing reliance on real patient samples while preserving downstream performance. Using guidance from histological prototypes during sampling, our approach ensures biologically and diagnostically meaningful variations in the generated data. We demonstrate that self-supervised features trained on our synthetic dataset achieve competitive performance despite using ~60x-760x less data than models trained on large real-world datasets. Notably, models trained using our synthetic data showed statistically comparable or better performance across multiple evaluation metrics and tasks, even when compared to models trained on orders of magnitude larger datasets. Our hybrid approach, combining synthetic and real data, further enhanced performance, achieving top results in several evaluations. These findings underscore the potential of generative AI to create compelling training data for digital pathology, significantly reducing the reliance on extensive clinical datasets and highlighting the efficiency of our approach.

Figures (3)

• Figure 1: Overview of the proposed approach. A. A WSI is segmented and patched into a set of non-overlapping patches. A compressed feature for each patch is obtained through a pre-trained feature encoder. K-Means clustering is performed to identify prototypes within each cancer type. B. A latent autoencoder (AE) and a latent diffusion model (LDM) are trained on a large-scale dataset of histopathology images paired with prototype values obtained from clustering for conditional image synthesis under the guidance of a trained latent classifier. C. Sampling a fixed number of images from the LDM, guided by each prototype, to construct a synthetic dataset for SSL model training. D. We test the proposed method and baselines with few-shot learning on clinical downstream tasks (subtyping and survival prediction).
• Figure 2: Example of a WSI from TCGA-UCS (Uterine Carcinosarcoma) with its corresponding prototype map, showing 18 detected clusters for this cancer type, along with prototype distribution for the slide and patch examples from the four largest clusters.
• Figure 3: Comparison of real images from the training subset with images generated using prototype guidance for three prototypes, each representing a different tissue type: DLBC (Diffuse Large B-Cell Lymphoma), ACC (Adrenocortical Carcinoma), and KIRP (Kidney Renal Papillary Cell Carcinoma).